Adversarial Augmentation for Robust Speech Separation

Speech separation is the task of separating single signals from the given mixture of multiple speakers. Neural networks trained for speech separation usually work well on artificial data but they often fail on real-world examples. To improve their behavior on real-world mixtures it is possible to use training data augmentations such as noise addition. Nevertheless, the power of these augmentations is limited as they have to be manually designed.     In this work, the modified version of the generative adversarial networks (GAN) model could improve this process by generating augmentations depending on the separation performance on these data. Speech separation could be then made more robust with each generator and separator training step. This system was subjected to experimentation. During these experiments, the parameters have been tuned to find the best setting that will successfully train the GAN model without collapsing. This setting was found and the most robust model from the training was selected and evaluated. Results show that the separator model trained by the GAN model does not achieve any significant improvement from the original separator model pretrained on the WSJ0-2mix dataset during the testing on the WHAM dataset. Nevertheless, another evaluation shows that the separator model trained by the GAN model is significantly more robust than the original one towards the generated noises.

speech separation, GAN, adversarial augmentations, robust neural network